Ground time sequence observation image wake cloud identification method and system

Abstract

The invention discloses a ground time sequence observation image wake cloud identification method and system, and belongs to the technical field of image processing, and the method comprises the steps: obtaining a video frame image of a current wake cloud, and the corresponding shooting time and coordinate information; analyzing a time label of the video frame image according to the video frame image and the shooting time, and extracting a time feature map of the video frame image by constructing a time feature vector; obtaining channel attention and space attention of the video frame image according to the video frame image and the coordinate information, and extracting a space feature map of the video frame image; according to the time feature map and the space feature map, performing channel dimension splicing on the time feature tensor and the space feature tensor to generate a combined feature map; and carrying out classification identification on the combined feature map to obtain the category of the wake cloud. According to the method, accurate identification and classification of wake clouds from generation to diffusion stages can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and more particularly to a ground time sequence observation image contrail recognition method and system. BACKGROUND

[0002] Contrail is a fine ice crystal cloud band formed by water vapor emitted by high-altitude aircraft during flight and rapidly condensed in a low-temperature environment, usually having a clear linear form, and is one of the common visible features of aircraft operation at high altitudes. Contrail not only affects the stealth of the flight path of the aircraft, but also may interfere with the results of meteorological monitoring.

[0003] Currently, the monitoring means for contrail mainly includes two kinds: one is remote sensing recognition based on high-resolution satellite images, and the other is fixed-point observation based on ground visible light cameras. Whether it is a remote sensing image or a ground image, the core of the processing is to automatically recognize and classify the contrail by means of image recognition algorithm in computer vision. The current mainstream method mainly adopts deep learning technology, including ResNet, YOLO, UNet and other target detection and segmentation frameworks. These methods have good performance in traditional cloud recognition, smoke detection, trajectory extraction and other scenes, and are therefore widely tried in the automatic recognition task of contrail. In practical application, researchers usually construct an image dataset containing contrail, and train and test based on the above-mentioned models, so as to realize the detection and segmentation of the contrail target.

[0004] Although the prior art has achieved automatic identification of contrail clouds to some extent, there are still the following technical defects and limitations, including the following points: lack of modeling ability of the time sequence characteristics of contrail clouds. The morphology of contrail clouds will change significantly over time: in the early stage, it has the characteristics of being slender and clear in edge; in the middle stage, it will gradually spread and the shape will become wide; in the late stage, it may even be integrated with natural clouds and become difficult to distinguish. Most of the current mainstream recognition algorithms are trained based on the spatial features of static images, and lack the modeling ability of time sequence labels and evolution state, which leads to a significant decrease in accuracy when the model identifies early and late contrail clouds. The existing data set is not timely and authentic. Most of the public contrail image data sets are derived from aerial remote sensing or simulated images, although the image clarity is high, but lack of real time label, and the environmental conditions are relatively single, which cannot reflect the complex light, background cloud interference and other problems under the real ground observation conditions. This makes the model have poor generalization ability when deployed in the field, and the recognition stability is insufficient. Light and weather interference has a significant impact on recognition performance. The ground collected images are greatly affected by the sun angle, cloud brightness, background sky light and other factors. Traditional image recognition models are prone to misidentifying contrail clouds as natural clouds, aircraft tracks or other linear image elements under low contrast or high dynamic light conditions, reducing the recognition accuracy. Lack of end-to-end acquisition-labeling-recognition system. Current researches mainly focus on model algorithm improvement, but rarely involve optimization of complete data acquisition and labeling process, lack of overall solution for time sequence evolution and low-cost deployment, which limits the application and real-time performance of this technology.

[0005] In summary, since the morphology of contrail clouds changes significantly over time, the existing technology cannot reflect the complex light, background cloud interference under the real ground observation conditions when the image is acquired, which makes the model have poor generalization ability when deployed in the field, thereby affecting the recognition accuracy. SUMMARY

[0006] In view of the problems existing in the above-mentioned field, the present application provides a ground time sequence observation image contrail cloud recognition method and system, which acquires image information with time labels, extracts time feature maps by constructing time feature vectors in the feature extraction stage, to improve the time perception ability of the recognition model; generates weight maps of channel attention and spatial attention features through global average pooling and maximum pooling operations, and enhances the middle features of the weight maps in the channel dimension and the spatial dimension respectively, to extract spatial feature maps, to improve the attention of the model to the main area of the contrail cloud, and to fuse the time feature maps and the spatial feature maps, to realize accurate recognition and classification of the contrail cloud from the generation to the diffusion stage.

[0007] To solve the above technical problems, the present application discloses a ground time sequence observation image contrail cloud recognition method, comprising the following steps: Acquire the video frame image of the current contrail cloud, and its corresponding shooting time and coordinate information; According to the video frame image and the shooting time, the time feature map of the contrail cloud video frame image is extracted by constructing a time feature vector; According to the video frame image and the coordinate information, the channel attention and the spatial attention of the video frame image are obtained, the weight map of the channel attention and the spatial attention features is generated through global average pooling and maximum pooling operation, and the middle features of the weight map are enhanced in the channel dimension and the spatial dimension respectively, and the spatial feature map of the video frame image is extracted; According to the time feature map and the spatial feature map, the time feature and the spatial feature tensor are spliced in the channel dimension to generate a joint feature map; The joint feature map is classified and recognized to obtain the category of the contrail cloud.

[0008] Preferably, the category of the contrail cloud is recognized by an improved YOLOV8 network model, and the improved YOLOV8 network model specifically comprises: taking a YOLOv8-tiny model as a detection skeleton structure, including a Backbone, a Neck, a Head and a time sequence label module; An attention mechanism CBAM is inserted in the c2f of the previous feature extraction layer and the Conv of the next feature extraction layer of the Backbone, which is used to apply channel attention and spatial attention to the intermediate feature map, and the spatial features are extracted through joint modeling of the channel attention and the spatial attention, guiding the automatic focusing on the contrail cloud area in the feature extraction stage; The time sequence label module is used to define the image file name according to the video frame image and the shooting time, parse the time label through the image file name, and convert the time category information into a discrete vector through one-hot encoding, and then map it to a 64-dimensional time feature vector through a fully connected layer; The 64-dimensional time feature vector is fused with the spatial features in the Neck through a concat mode to generate a joint feature map; The joint feature map is classified and recognized by the Head.

[0009] Preferably, the improved YOLOV8 network model introduces a time consistency loss function during training, and adds a similarity constraint term between the prediction results of the same target at different time frames on the basis of the original YOLO loss in the Head of the YOLOv8-tiny model; The similarity constraint term is: ; Wherein, represents the prediction probability vector of the same target at different times; a time-adjacent weight function, is an adjustment factor.

[0010] Preferably, it further comprises: According to the shooting time and the state of the contrail cloud of each frame of image obtained, a one-to-one correspondence is established, and a contrail cloud evolution state label system is constructed; According to the contrail cloud evolution state label system and the coordinate information corresponding to each image, the target classification of the video frame image is carried out, and each type of labeled image is obtained.

[0011] The construction of the contrail cloud evolution state label system specifically includes: By analyzing the morphological evolution process of the contrail cloud, the target is divided into three states of just generated, corresponding to 0min; generated about thirty minutes, corresponding to 30min; generated about one hour, corresponding to 60min, combined with the shooting time, and a non-contrail cloud category is added as a control class.

[0012] Preferably, the classification and identification of the joint feature map further includes outputting the target position, evolution state category and confidence information of the contrail cloud.

[0013] Preferably, the acquisition of the video frame image of the current contrail cloud specifically includes: A plurality of point positions are selected in the shooting area, a fixed elevation angle is set for the camera, and the video frame image of the contrail cloud in the current state is obtained by timed shooting; The video frame image of the contrail cloud is detected through edge computing, the linear structure feature in the image is detected, the invalid image without significant high-altitude track is excluded, the image qualified in the quality preliminary screening is obtained, and the image qualified in the quality preliminary screening is labeled.

[0014] Preferably, the coordinate information is GPS positioning coordinates.

[0015] Preferably, it further includes a ground time sequence observation image contrail cloud identification system, which comprises: An image acquisition module is used to acquire the video frame image of the current contrail cloud, and the corresponding shooting time and coordinate information thereof; A feature extraction module is used to extract the time feature map of the contrail cloud video frame image by constructing a time feature vector according to the video frame image and the shooting time; according to the video frame image and the coordinate information, the channel attention and the spatial attention of the video frame image are obtained, the weight map of the channel attention and the spatial attention feature is generated through global average pooling and maximum pooling operation, and the middle feature of the weight map is enhanced in the channel dimension and the spatial dimension respectively, and the spatial feature map of the video frame image is extracted; A feature fusion module is used to generate a joint feature map by channel dimension splicing of the time feature and the spatial feature tensor according to the time feature map and the spatial feature map; The identification module is used to classify and identify the joint feature map to obtain the category of the wake cloud.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention proposes a ground-based temporal observation image contrail cloud recognition method. This method acquires time-labeled contrail cloud images and, during the feature extraction stage, constructs a temporal feature vector to extract a temporal feature map, thereby enhancing the temporal perception capability of the recognition. Based on video frame images and coordinate information, the method obtains the channel attention and spatial attention of the video frame images. Global average pooling and max pooling operations are used to generate weight maps of the channel attention and spatial attention features. The intermediate features of the weight maps are enhanced in both the channel and spatial dimensions to extract the spatial feature map of the video frame images. This improves the model's focus on the main region of the contrail cloud, suppresses background interference, and enhances the stability of the recognition boundary, maintaining excellent performance even under complex conditions such as strong light and cluttered clouds. The temporal and spatial feature tensors are concatenated along the channel dimension for feature map fusion, generating a joint feature map. This joint feature map not only possesses spatial perception capability but also identifies state differences of the contrail cloud in the temporal evolution dimension. Classification and recognition of the joint feature map enable accurate identification and classification of contrail clouds from their generation to diffusion stages. Attached Figure Description

[0017] Figure 1 This is a flowchart of the ground time-series observation image wake cloud recognition method proposed in this invention; Figure 2 The improved YOLOv8 network model architecture constructed for this invention; Figure 3 This is a schematic diagram of the overall structure of the identification system of the present invention; Figure 4 The identification results provided in the embodiments of the present invention. Detailed Implementation

[0018] The following will refer to the appendices in the embodiments of the present invention. Figures 1-4 The technical solutions in the embodiments of the present invention will be clearly and completely described. It should be understood that the terminology used in the present invention is only for describing particular implementation methods and is not intended to limit the present invention.

[0019] like Figure 1 As shown, this invention proposes a method for identifying contrails in ground-based time-series observation images to improve the accuracy and stability of contrails identification at different stages. The method specifically includes the following steps: Acquire the current video frame image of the contrail cloud, along with its corresponding capture time and coordinate information (GPS positioning coordinates); According to the video frame image and the shooting time, the time label of the video frame image is analyzed, the time feature vector is constructed, and the time feature map of the video frame image is extracted; According to the video frame image and the coordinate information, the channel attention and spatial attention features are extracted, the weight map is generated through global average pooling and maximum pooling operations, and the middle features of the weight map are enhanced in the channel dimension and the spatial dimension, respectively, to extract the spatial feature map of the video frame image. According to the time feature map and the spatial feature map, the time feature and the spatial feature tensor are spliced in the channel dimension to generate a joint feature map. The joint feature map is classified and recognized to obtain the category of the wake cloud.

[0020] Specifically, the classification and recognition of the joint feature map also includes outputting the target position, evolution state category and confidence information of the wake cloud.

[0021] The category of the wake cloud is recognized by an improved YOLOV8 network model, which specifically includes: taking a YOLOv8-tiny model as a detection skeleton structure, including Backbone, Neck, Head and a time sequence label module. The attention mechanism CBAM is inserted into the c2f of the previous feature extraction layer and the Conv of the next feature extraction layer of Backbone, which is used to apply channel attention and spatial attention to the intermediate feature map, extract channel attention and spatial attention features of the wake cloud video frame image, and perform spatial feature extraction through joint modeling of channel attention and spatial attention, guiding automatic focusing on the wake cloud area in the feature extraction stage, and improving the recognition ability in complex background.

[0022] The lightweight neural network model is combined with the attention mechanism to enhance the robustness of complex environment recognition. Through the CBAM module, the model's attention to the wake cloud stem area is improved, background interference is suppressed, recognition boundary stability is improved, and excellent performance is still achieved under strong light, miscellaneous clouds and other complex conditions.

[0023] The time sequence label module (introducing a time sequence label mechanism) is used to define the image file name according to the video frame image and the shooting time, parse the time label through the image file name, convert the time category information into a discrete vector through one-hot encoding, map it to a 64-dimensional time feature vector through a fully connected layer, and fuse the 64-dimensional time feature vector with the spatial feature in the Neck through the concat method to generate a joint feature map, realize context modeling of the time sequence evolution state, and classify and recognize the joint feature map through the Head.

[0024] The improved YOLOV8 network model can improve the time perception ability of the recognition model in training by introducing a time label mechanism.

[0025] That is, on the basis of the original YOLO loss in the Head of the YOLOv8-tiny model, a similarity constraint term between the prediction results of the same target at different time frames is added, which is used for classification and recognition of the joint feature map.

[0026] Specifically, the similarity constraint term is specifically: ; Among them, represents the prediction probability vector of the same target at different times; represents a time adjacency weight function, is an adjustment factor.

[0027] By establishing a one-to-one correspondence between the image shooting time and the tail cloud state, the time state is taken as one of the model input features, which significantly improves the classification accuracy of tail clouds at different generation stages, and solves the problem that the traditional model cannot distinguish between "nascent tail clouds" and "diffusion tail clouds".

[0028] It also includes establishing a one-to-one correspondence according to the shooting time and the tail cloud state of each frame of the image of the tail cloud obtained, and constructing a tail cloud evolution state label system; According to the tail cloud evolution state label system and the coordinate information corresponding to each image, the target classification of the video frame image is carried out, and each type of label image is obtained.

[0029] The tail cloud evolution state label system is constructed, specifically including: By analyzing the morphological evolution process of the tail cloud, the target is divided into three states of just generated, corresponding to 0min; generated for about thirty minutes, corresponding to 30min; generated for about one hour, corresponding to 60min, and a non-tail cloud category is added as a control class.

[0030] The tail cloud video frame image under the current state is obtained, specifically including: A plurality of point positions are selected in the shooting area, a fixed elevation angle is set for the camera, and the video frame image of the tail cloud under the current state is obtained by timing shooting; The video frame image of the tail cloud is detected through edge computing, the linear structure feature in the image is detected, the invalid image without obvious high-altitude track is excluded, the image qualified in quality preliminary screening is obtained, and the image qualified in quality preliminary screening is labeled.

[0031] The application also provides a ground time sequence observation image tail cloud recognition system, which includes: An image acquisition module (image acquisition module) is configured to acquire a video frame image of a current contrail cloud, and corresponding shooting time and coordinate information.

[0032] Raspberry Pi 4B devices are deployed at 5-10 typical points (high-rise buildings, open spaces) in the shooting area, each equipped with an HQ Camera Module (12 million pixels) and a GPS coordinate synchronization module as shown in Figure 3 for fixing the shooting airspace image. The camera is directed north and has a fixed elevation angle to avoid misphotographing the landscape. A frame of image is photographed every 15 seconds, and the collection time is not less than 30 days. The system runs all day long, and the daily average image quantity of a single point is about 5760. A total of about 1728000 images (30 days) are collected at 10 points. The video frame image of the contrail cloud is detected by edge computing to detect the linear structure features in the image, exclude invalid images without significant high-altitude tracks, obtain images that pass the quality preliminary screening, and label the images that pass the quality preliminary screening. After excluding images without contrail clouds, about 10,000 to 20,000 contrail cloud-related images are retained as candidate samples.

[0033] The images are stored in JPEG format, and the images are named with timestamps and the shooting coordinates are recorded synchronously (for labeling reference and backtracking analysis).

[0034] A feature extraction module is configured to parse the time label of the video frame image of the contrail cloud according to the video frame image of the contrail cloud and the shooting time, construct a time feature vector, and extract a time feature map of the video frame image of the contrail cloud; acquire channel attention and spatial attention of the video frame image according to the video frame image of the contrail cloud and the coordinate information, generate a weight map of the channel attention and the spatial attention feature through global average pooling and maximum pooling operations, and enhance the intermediate features of the weight map in the channel dimension and the spatial dimension, respectively, to extract a spatial feature map of the video frame image.

[0035] Specifically, the time feature map and the spatial feature map of the contrail cloud are extracted by an improved YOLOV8 network model.

[0036] As shown in Figure 2 , the improved YOLOV8 network model takes YOLOv8-tiny model as the detection skeleton structure, wherein the YOLOv8-tiny model includes: Backbone (CSPDarknet-Tiny, with residual connection and C2f structure), Neck (PAN-FPN structure, for multi-scale feature fusion), and Detection Head (three-layer output branch, responsible for small, medium and large scale object detection, respectively).

[0037] The improved YOLOV8 network model inserts an attention mechanism (CBAM module) in the c2f of the previous feature extraction layer and the Conv of the subsequent feature extraction layer of the Backbone, obtains the channel attention and spatial attention of the video frame image of the contrail cloud, combines the channel attention and spatial attention, applies the channel attention and spatial attention to the intermediate feature map, performs spatial feature extraction through the combined modeling of the channel attention and spatial attention, generates a weight map of the channel attention and spatial attention features through global average pooling and maximum pooling operations, and respectively enhances the intermediate features of the weight map in the channel dimension and the spatial dimension, extracts the spatial feature map of the image, guides the model to focus on the contrail cloud region, and improves the recognition ability in a complex background.

[0038] A time sequence label module is introduced, i.e., a time sequence label mechanism is introduced, an image file name is defined according to a video frame image and a shooting time, a time period label (0, 30 min, 60 min) is parsed from the image file name, one-hot encoding is constructed as a time vector, a linear mapping is performed to obtain a 64-dimensional time feature vector, and the time feature vector is fused with image spatial features in a Neck layer through a concat mode to realize context modeling of a time sequence evolution state.

[0039] A time consistency loss function is introduced: a time consistency constraint term is added to the original YOLO loss of the Head .

[0040] Further comprising: According to the shooting time and the contrail cloud state of the image of each frame of the contrail cloud, a one-to-one correspondence is established, and a contrail cloud evolution state label system is constructed; According to the contrail cloud evolution state label system and the coordinate information corresponding to each image, target classification is performed on the video frame image to obtain each type of labeled image.

[0041] In particular, the candidate samples are uploaded to the CVAT platform, and the labeling personnel perform artificial fine labeling by combining time and image features.

[0042] The contrail cloud evolution state label system is constructed, and the classification is as follows: Class 0: Contrail cloud-just generated (thin and clear lines, consistent with the flight path); Class 1: Contrail cloud-generated for 30 minutes (lines have spread, outline is blurred); Class 2: Contrail cloud-generated for 1 hour (greatly spread, difficult to distinguish from natural clouds); Class 3: Non-contrail cloud (natural cloud, blue sky background, other cirrus clouds, etc.).

[0043] Each image supports multi-target labeling, and the labeling format adopts the YOLO standard: class x_center y_center width height, and all coordinate information is normalized to the image resolution (default 1920*1080). The total number of labeled samples is controlled to be not less than 10,000 images, and the number of images of each label is not less than 2,000.

[0044] The feature fusion module is configured to perform channel dimension splicing on the time feature and the space feature tensor according to the time feature map and the space feature map, and generate a joint feature map. The recognition module is configured to perform classification recognition on the joint feature map, and obtain a category of the wake cloud.

[0045] The obtained video frame images are divided into a training set, a verification set and a test set according to a 7:2:1 ratio; the number of training set images is controlled to be about 7,000; the number of verification set images is 2,000, and the number of test set images is 1,000, so as to ensure the balance of each label.

[0046] The training parameter setting includes: Framework: PyTorch; Optimizer: SGD + Cosine Annealing; Learning rate: 0.001 initially; Batch Size: 16; Epochs: 150; Transfer learning: based on YOLOv8 pre-training weight fine-tuning; Enhancement strategy: using Mosaic, Mixup, random brightness, blur and other image enhancement methods.

[0047] Evaluation index: Using mAP 0.5 (mean Average Precision).

[0048] The system integration module of the application is composed of the following sub-modules: Image acquisition terminal: Raspberry Pi device carries image acquisition service, runs Python program for regular acquisition, preliminary screening and image uploading, supports local caching and network retransmission mechanism.

[0049] Recognition inference module: the backend server deploys the improved YOLOv8-tiny model, uses ONNX+TensorRT for acceleration deployment, and supports 30 FPS real-time recognition.

[0050] Control and visualization interface: Flask+Vue is used to build the control interface, which supports image preview, recognition result display, wake cloud state statistics and parameter adjustment.

[0051] Data feedback and management: The recognition results are written to the local SQLite database in real time and uploaded to the remote platform synchronously, supporting functions such as trace cloud quantity statistics, generation time distribution analysis, and dynamic evolution trend visualization.

[0052] Example To better demonstrate the contrail cloud recognition method proposed in this invention, this embodiment employs a multi-point ground deployment approach, setting up image acquisition terminals at typical airspace observation locations. Each terminal consists of a Raspberry Pi 4B embedded device, equipped with an HQ Camera Module image acquisition module and a GPS positioning module. The image acquisition terminal is pre-installed with image control software, capable of automatically performing airspace shooting tasks in all weather conditions at a preset frequency, ensuring the acquisition of one high-definition image frame every 15 seconds. During the acquisition process, the camera is fixed facing north and maintains a specific elevation angle to avoid interference from ground objects. The system simultaneously records the image generation time and geographic coordinate information. All images are stored in JPEG format and named using timestamps to ensure the integrity and traceability of the image sequence.

[0053] After image acquisition, the device's local program automatically performs initial image quality screening, eliminating nighttime images, severely overexposed images, or invalid images, thereby reducing the burden on the backend recognition. The edge computing module determines whether there are obvious linear structures or high-altitude trajectory features in the images. Image data that passes the initial screening is uploaded to the server and stored in the database via a pre-defined network protocol. The system has a built-in breakpoint resume and caching mechanism, which can automatically resume uploading in the event of a network interruption, ensuring data integrity.

[0054] Uploaded image data is uniformly imported into the CVAT annotation platform, where annotators perform manual annotation based on image content and capture time information. By analyzing the morphological evolution of contrails after generation, annotators categorize image targets into three contrail evolution states: immediately after generation, approximately 30 minutes after generation, and approximately 1 hour after generation. A non-contrail category is also added as a control class. The bounding boxes and category information of targets in each image are uniformly annotated according to the YOLO format, and the coordinates are normalized to adapt to the requirements of neural network training. The number of annotated samples is controlled to be over 10,000 to ensure a balanced distribution of labels for each class, facilitating the construction of a high-quality, time-series-defined training dataset.

[0055] The tail cloud recognition system provided by the embodiment as a whole comprises an image acquisition module, an image labeling module, an algorithm improvement module, a model training and testing module, and a system integration module. The system integration module comprises an identification inference module, a control and visualization interface, and a data return and management and dynamic evolution trend analysis module. The image acquisition module realizes a timing acquisition function based on a Raspberry Pi device running a Python program, reads NMEA data of a GPS module through a serial port to realize coordinate synchronization, binds a timestamp and longitude and latitude to each frame of image and stores them in a local cache queue, adopts an LRU strategy to manage a cache space with a capacity of 100 frames, and starts an exponential backoff retransmission mechanism to preferentially send the latest data when the network is interrupted. The identification inference module deploys an improved YOLOv8 network model, takes YOLOv8-tiny as a backbone network, adds a CBAM attention mechanism at the end of the backbone network to enhance the tail cloud feature extraction capability, introduces a time feature fusion module to process continuous 3 frames of feature maps through an LSTM unit to capture dynamic changes, and adds a time consistency loss function in the loss function to make the identification results of adjacent frames change suddenly; after the improved YOLOv8 network model is converted from PyTorch to ONNX, FP16 quantization and layer fusion optimization are performed through a TensorRT engine, and 30FPS real-time inference with a delay of ≤33ms is realized.

[0056] The control and visualization interface is built by using Flask+Vue, and four-level classification standards of tail cloud states are defined: 0 level represents no cloud, 1 level corresponds to a single block with an area less than 50% of the image area, 2 level is a single block with an area greater than 50%, and 3 level represents multiple tail clouds existing at the same time. The interface integrates a dynamic evolution trend analysis module, applies an exponential smoothing method based on SQLite time series data to predict the number of tail clouds, generates a tail diffusion thermodynamic map by using an OpenGL engine, and supports 1-hour sliding window analysis. A double- SQLite database architecture is adopted for the data return and management system, the terminal database caches original images and GPS metadata and automatically cleans old data 24 hours ago, and the platform database permanently stores identification results and statistical indicators; after the network is restored, the incremental synchronization of cached data during the network interruption is realized through an HTTPS protocol.

[0057] A 7:2:1 data set division ratio is adopted for the model training end, a Mosaic+MixUp combination enhancement strategy is applied, random HSV jitter and ±15-degree affine transformation are used to improve data diversity. Migration learning and parameter optimization techniques are introduced in the training process, and the evaluation system comprises core indicators such as mAP@0.5 / 0.95, recall rate, precision rate and inference speed, to ensure that the model meets the performance requirements in actual deployment. The modules realize collaborative work through standardized interfaces, forming a complete closed-loop system from data acquisition, real-time analysis to visualization presentation.

[0058] This embodiment takes YOLOv8-tiny model as the skeleton recognition network model, and makes targeted improvements on the network structure.

[0059] Firstly, the attention mechanism (CBAM module) is introduced in the backbone network part. Through the joint modeling of channel attention and spatial attention, the neural network is guided to automatically focus on the contrail cloud area in the feature extraction stage, and the influence of interference factors such as natural clouds and sunlight scattering in the background is weakened. The module generates a weight map through global average pooling and maximum pooling operations after the feature map output, and enhances the middle features of the weight map in the channel dimension and the spatial dimension respectively, thereby improving the sensitivity of the model to the contrail cloud area and the boundary judgment accuracy.

[0060] Secondly, the time label module is introduced in the feature extraction stage. The system parses the time label through the time label mechanism according to the image file name, and converts the time category information into a discrete vector through one-hot encoding, and then maps it into a 64-dimensional time feature vector through a fully connected layer.

[0061] In the feature fusion stage, the 64-dimensional time feature vector and the image spatial feature tensor are spliced in the channel dimension through the concat method in the Neck layer, generating a joint feature map, so that the model not only has spatial perception ability, but also can identify the state difference of the contrail cloud in the time evolution dimension.

[0062] To enhance the recognition consistency of the model in the time dimension, a time consistency loss function is also introduced. The loss function adds a similarity constraint term between the prediction results of the same target at different time frames based on the original target detection loss, thereby encouraging the model to maintain semantic coherence and stage rationality in the prediction results of the contrail cloud evolution process.

[0063] In the model training stage, the PyTorch framework is used to build the neural network, and the pre-trained parameters of YOLOv8 are loaded in a transfer learning manner. The data set is divided into training set, validation set and test set to ensure the generalization ability of the model. During the training process, the system uses Mosaic and Mixup enhancement, brightness disturbance, and blur and other image enhancement methods to improve the robustness of the model under complex lighting and blurred cloud conditions. The loss function is composed of boundary box error, class cross entropy and time consistency loss, and the cosine annealing strategy is used to adjust the learning rate to improve the convergence efficiency.

[0064] After training, the model is exported in ONNX format and deployed on a backend server, using the TensorRT inference engine to achieve high concurrency and high-speed real-time recognition inference services. The server provides a RESTful interface for the front end to call, which returns the contrail cloud recognition results immediately after the image is uploaded, including the target position, evolution state category and confidence information. The recognition results are written to the local database in real time and transmitted to the remote center platform for subsequent statistical analysis and trend modeling.

[0065] The system integration part uses Flask as the backend service framework and combines Vue to realize data visualization interface. Users can view the image acquisition situation, recognition results and contrail cloud quantity statistics in real time on the interface. The system supports dynamic parameter configuration and model updating, and provides visualization modules for contrail cloud generation time distribution and state evolution trend, which facilitates researchers to conduct long-term monitoring and meteorological correlation analysis.

[0066] The overall system has technical advantages such as low-cost deployment, high time sequence sensitivity and strong environmental adaptability, and is suitable for contrail cloud recognition tasks in complex weather backgrounds, and can be widely used in aviation stealth evaluation, airspace monitoring and warning, and meteorological data correction. The invention has significant technical innovation points in hardware deployment methods, image acquisition strategies, time sequence label embedding mechanisms and time consistency loss design, effectively overcoming the technical bottlenecks of traditional methods in contrail evolution recognition, and has good practicality and scalability.

[0067] As shown in Figure 4 , the typical output results of the improved YOLOV8 network model after recognizing the video frame image samples collected by the embodiment are shown. The figure clearly shows the detection frame of the improved YOLOV8 network model for different targets in the image and the corresponding class label and confidence. Specifically:

[0068] Figure 4 The main targets identified in the figure are two types of contrail clouds, Contrail=1hour and Contrail=Fresh, where: Contrail=1hour is identified multiple times with confidence of 0.90, 0.80, 0.52, 0.31, indicating that this target exists significantly in the scene and the improved YOLOV8 network model has high reliability in identifying it.

[0069] Contrail=Fresh appears most frequently with confidence of 0.80, 0.57, 0.82, 0.62, 0.69, 0.45, and is the main recognition category in the image, but its confidence fluctuation range is relatively large, reflecting the differences in target state or imaging conditions.

[0070] Figure 4 The recognition results also include other significant targets, among which Sun is recognized with a high confidence of 0.90, indicating that there is a significant feature of the sun in the image. Parasite is recognized with the highest confidence of 0.93, indicating that the improved YOLOV8 network model has strong recognition ability for such features and the target is obvious in the image. Unknown is a target detected by the improved YOLOV8 network model, and the corresponding confidence is 0.62, but the specific category cannot be determined.

[0071] In summary, Figure 4 The tail cloud recognition results shown intuitively show that the improved YOLOV8 network model can effectively identify multiple key targets in the video frame image of the tail cloud, especially for Contrail=1hour, Contrail=Fresh, Sun and Parasite categories. At the same time, the results also contain the detection box of the unknown category, reflecting the challenges in the actual recognition scene and the confidence level difference of the improved YOLOV8 network model. At the same time, Figure 4 The density of the target box in the image also implies that the tail cloud and related targets in the image are relatively concentrated.

[0072] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can make equivalent replacements or changes to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

[0073] In addition, unless otherwise specified, all technical and scientific terms used in the present application have the same meaning as generally understood by those skilled in the art. All documents mentioned in the specification are incorporated by reference to disclose and describe the methods related to the documents. In the event of any conflict with any incorporated document, the content of the specification shall prevail.

Claims

1. A ground time-series observation image contrail recognition method, characterized by, The method comprises the following steps: obtaining a video frame image of a current contrail cloud, and corresponding shooting time and coordinate information thereof; according to the video frame image and the shooting time, constructing a time feature vector to extract a time feature map of the video frame image of the contrail cloud; according to the video frame image and the coordinate information, obtaining channel attention and spatial attention of the video frame image, generating a weight map of the channel attention and the spatial attention feature through global average pooling and maximum pooling operations, and enhancing the middle feature of the weight map in the channel dimension and the spatial dimension respectively to extract a spatial feature map of the video frame image; splicing the time feature and the spatial feature tensor in the channel dimension to generate a joint feature map; performing classification recognition on the joint feature map to obtain the category of the contrail cloud.

2. The ground time-series observation image contrail identification method of claim 1, wherein, The category of the contrail cloud is recognized by an improved YOLOV8 network model, and the improved YOLOV8 network model specifically comprises: taking a YOLOv8-tiny model as a detection skeleton structure, including a Backbone, a Neck, a Head and a time sequence label module; CBAM is inserted into c2f of a front feature extraction layer and Conv of a rear feature extraction layer of the Backbone to apply channel attention and spatial attention to the middle feature map, to guide automatic focusing on the contrail cloud region in the feature extraction stage through joint modeling of the channel attention and the spatial attention for spatial feature extraction; the time sequence label module is used to define an image file name according to the video frame image and the shooting time, parse the time label through the image file name, convert the time category information into a discrete vector through one-hot encoding, and then map the discrete vector into a 64-dimensional time feature vector through a fully connected layer; the 64-dimensional time feature vector is fused with the spatial feature through concat in the Neck to generate a joint feature map; the joint feature map is classified and recognized by the Head.

3. The ground time-series observation image contrail identification method of claim 2, wherein, When training the improved YOLOV8 network model, a time consistency loss function is introduced, and a similarity constraint term between prediction results of the same target at different time frames is added to the original YOLO loss in the Head of the YOLOv8-tiny model; the similarity constraint term is: ； wherein, denotes the predicted probability vector of the same target at different times; denotes a time adjacency weight function, is an adjustment factor.

4. The ground time-series observation image contrail identification method of claim 3, wherein, further comprising: establishing a one-to-one correspondence relationship according to the shooting time and the contrail cloud state of each frame of the image of the contrail cloud, and constructing a contrail cloud evolution state label system; performing target classification on the video frame image according to the contrail cloud evolution state label system and the coordinate information corresponding to each image to obtain each type of label image.

5. The ground time-series observation image contrail identification method of claim 4, wherein, The construction of the contrail cloud evolution state label system specifically comprises: by analyzing the morphological evolution process of the contrail cloud, the target is divided into three states of just generated, corresponding to 0min; generated for about thirty minutes, corresponding to 30min; generated for about one hour, corresponding to 60min, and a non-contrail cloud category is added as a control class.

6. The ground time-series observation image contrail identification method of claim 5, wherein, The classification recognition of the joint feature map further comprises outputting the target position, the evolution state category and the confidence information of the contrail cloud.

7. The ground time-series observation image contrail identification method of claim 1, wherein, The video frame image of the current contrail cloud is acquired, and specifically includes the following steps: A plurality of point positions are selected in a shooting area, a fixed elevation angle is set for the camera, and a video frame image of the contrail cloud in a current state is acquired by timing shooting; The video frame image of the contrail cloud is subjected to edge calculation, linear structure features in the image are detected, invalid images without significant high-altitude tracks are excluded, images that pass the quality preliminary screening are obtained, and the images that pass the quality preliminary screening are labeled.

8. The ground time-series observation image contrail identification method of claim 1, wherein, The coordinate information is a GPS positioning coordinate.

9. A ground time-series observation image contrail recognition system, characterized by, It includes: An image acquisition module is configured to acquire a video frame image of a current contrail cloud, and corresponding shooting time and coordinate information thereof; A feature extraction module is configured to extract a time feature map of the video frame image of the contrail cloud by constructing a time feature vector according to the video frame image and the shooting time; acquire channel attention and spatial attention of the video frame image according to the video frame image and the coordinate information; generate a weight map of the channel attention and the spatial attention feature by global average pooling and maximum pooling operations; and enhance the middle feature of the weight map in the channel dimension and the spatial dimension, respectively, to extract a spatial feature map of the video frame image; A feature fusion module is configured to perform channel dimension splicing on a time feature and a spatial feature tensor according to the time feature map and the spatial feature map to generate a joint feature map; An identification module is configured to perform classification identification on the joint feature map to obtain a category of the contrail cloud.

Images